Does BRR collect more data?

No. It constructs weighted replicates from the observed sample.

Does it apply unchanged to every design?

No. Standard BRR has specific paired-PSU requirements, and other designs or variants need appropriate construction.

Is the replicate mean always the full-sample estimate?

Not necessarily for nonlinear statistics; follow the specified centering convention.

AI Glossary / Statistics and probability / Advanced

Balanced repeated replication

Definition

Balanced repeated replication, or BRR, is a method for estimating the sampling variance of survey statistics using systematically constructed replicate weights. In its standard form, it applies to a stratified design with two primary sampling units per stratum. Each replicate keeps one unit from each pair and adjusts its weight, allowing variation across replicate estimates to reflect the sample design.

What it means

A complex survey's observations are not generally equivalent to an independent simple random sample. Stratification, clustering, and unequal selection probabilities affect how much a statistic would vary across repeated samples.

BRR addresses that variability by recalculating the statistic under a balanced set of weight perturbations. The replicates are computational constructions from the observed sample, not newly collected independent surveys.

The primary sampling unit, or PSU, may be a cluster containing many respondents. Standard BRR perturbs whole PSUs within strata rather than randomly deleting individual respondents across the dataset. Analysts using published replicate weights must follow the provider's variance formula and scaling conventions.

How it works

First calculate the full-sample estimate using the survey's main weights. Construct a collection of half-samples by selecting one of the two PSUs in each stratum. In the basic construction, selected PSUs receive twice their original weights and unselected PSUs receive zero weight.

The pattern of selections is balanced, commonly using columns of a Hadamard matrix with orthogonal plus and minus signs. This arrangement lets a limited set of replicates represent the relevant stratum contrasts rather than requiring every possible half-sample.

Recalculate the entire statistic with each set of replicate weights. For ordinary BRR, a common variance estimator is the average squared difference between each replicate estimate and the full-sample estimate: sum of (theta_r − theta_full) squared, divided by the number of replicates R. The standard error is the square root of this variance estimate.

Variants need their own scaling. Fay's method retains contributions from both PSUs using less extreme perturbations and compensates through its variance factor. Real survey weights may also include nonresponse and calibration adjustments. Those adjustments, domain estimation, degrees of freedom, and unavailable replicate estimates must be handled according to the survey's specified procedure; the bare formula does not settle every implementation choice.

Three practical examples

A survey analyst estimates uncertainty in a weighted population mean using official BRR replicate weights.
A research team recalculates a ratio statistic across replicates instead of applying an independent-observation standard-error formula.
A statistical methods exercise compares the full-sample estimate with balanced half-sample estimates to show how PSU differences drive variance.

Think of it this way

Think of repeatedly viewing a survey through carefully arranged lenses that emphasize one member of each sampling pair. The lenses are designed to balance one another. How the estimate changes across those views helps measure sampling uncertainty, without pretending that each view is a separate new survey.

Worked example

Consider an original simplified design with two equally weighted strata and two PSUs per stratum. For teaching purposes, each PSU contributes one value. Stratum 1 has values 10 and 14; stratum 2 has values 20 and 28. All four full-sample weights are equal.

The full-sample mean is (10+14+20+28)/4 = 18. Use four balanced selection patterns: first-first, first-second, second-first, and second-second. Doubling the selected weights gives replicate means 15, 19, 17, and 21 respectively.

Their differences from 18 are −3, 1, −1, and 3. Ordinary BRR variance is (9+1+1+9)/4 = 5, so the standard error is sqrt(5), approximately 2.236. The unit is the original measurement unit for the standard error and its square for the variance.

The calculation uses division by R = 4, not the R−1 divisor of an ordinary sample variance. These replicates were constructed for a design-based variance estimator and are not four independent survey estimates. The example omits finite-population corrections and weight adjustments; it is an exact illustration of this specified BRR formula, not a universal recipe for all surveys.

Illustrative case study

Fictional case: Ashford Research receives a survey file containing a final weight and multiple BRR replicate weights. An initial dashboard treats all respondents as equally weighted and uses an ordinary independent-sample standard error.

The analysts rebuild the estimates using the main weight, then recompute each statistic with the supplied replicate weights and the provider's prescribed scaling. They retain the sample design when analyzing subgroups and document which statistics are estimable in every replicate.

The fictional correction produces a methodologically appropriate uncertainty workflow. It does not guarantee that sampling error captures nonresponse bias, measurement error, or every other source of uncertainty in the survey.

Common misconceptions and mistakes

Common mistakes include confusing PSUs with individual respondents, treating replicate estimates as independent samples, and substituting an R−1 divisor without justification.

Another error is using Fay-type replicate weights with an ordinary BRR scale factor. Survey-specific weighting and replicate construction matter. A large number of replicates also does not automatically create equally many independent degrees of freedom. Finally, filtering the dataset to a subgroup before applying a survey procedure can mishandle domain variance; use the design-aware method appropriate to the data and software.

Business use

BRR helps organizations report uncertainty for statistics from complex surveys, including means, proportions, ratios, and suitably supported model estimates. Published replicate weights can make design-aware analysis possible without exposing all details of the sampling structure.

Keep the provider's documentation with the data and record the exact variance convention. Check that the main estimate and replicate calculations use consistent definitions. Report standard errors or intervals with their assumptions, and distinguish sampling precision from the broader question of whether the survey accurately represents the intended population.

Check your understanding

Why is the example's variance 5 rather than 20/3?

Answer: The stated ordinary BRR formula averages squared deviations over four designed replicates; it is not the usual unbiased sample variance formula for independent observations.

Related terms

Stratified sampling
Standard error
survey weight
replicate weight
primary sampling unit
Hadamard matrix
Fay's method
design-based variance

Learn next

compare BRR with jackknife and survey bootstrap methods, study domain estimation, and examine how replicate scaling and design degrees of freedom affect intervals.

Degrees of freedom

Sources

Sources: SAS Institute, SURVEYMEANS documentation, Balanced Repeated Replication Method, https://support.sas.com/documentation/cdl/en/statug/66103/HTML/default/statug_surveymeans_details42.htm. The paired-PSU construction and ordinary variance convention were checked against the official documentation. The two-stratum example and fictional survey workflow are original illustrations.

Previous: Balanced accuracy score Next: Ball divergence